Apply guardrails to prevent the output of harmful content

Why it matters

Content filters are the first native control between a model and its users, and their thresholds decide what is blocked before any application logic runs.

What to practise

In the new Microsoft Foundry portal, deploy a model and send prompts through its default guardrail in the playground. Notice which refusals come from the model itself and which come from the filter. Then create a custom guardrail on the Guardrails page that sets the Hate, Violence, Sexual, and Self-harm risks to the highest blocking level for prompts and completions, apply it to the model deployment, and confirm it on the deployment’s details. The stricter guardrail may not change the earlier responses, because the default already blocks most lab-safe test content. The exercise notes that some of its technologies are in preview.

How it fits

Risk categories, severity thresholds, and where a guardrail is assigned are 4.2.1. The exercise applies the guardrail to a model, although the same step can select agents; for an agent, also check the agent overrides that 4.2.1 covers.